Forecasting what comes next in a sequence of measurements, such as daily sales or sensor readings, usually means training a separate model for each dataset. That is expensive and hard to maintain. Timestral offers a family of pretrained models that forecast many types of series without retraining. A new training step separates the act of predicting from the context used, so longer forecasts are learned without obscuring past data. The models also handle external information known in advance, such as holidays. On common benchmarks the largest model ranks first on every metric, and the next largest ranks second, beating much larger models. The weights are public, so teams can adopt accurate forecasting with far less effort.
Timestral models set new state of the art in time series forecasting
A family of four foundation models, from 5M to 209M parameters, achieves top performance on three major benchmarks through decoupled horizon forecasting and synthetic pretraining.
Independent
Adrien Petralia · Alexandre Gavaudan · Artjom Joosen · Etienne Goffinet · Richard Kurle · Yassine El Ouahidi · +350 more
Research Digest··3 min read
In one line
Timestral's four time series foundation models set new state-of-the-art accuracy on multiple forecasting benchmarks.
Before you read this
Three ideas this paper builds ontime series foundation modelA pretrained deep learning model for time series forecasting that generalizes across tasks or datasets.4 papers →synthetic data generationCreating training examples artificially, often using an LLM, to overcome real-data scarcity or privacy constraints.41 papers →role decouplingSeparating distinct functions such as planning, reasoning, and execution into separate components improves efficiency, quality, and reliability.39 papers →
Read it as
The authors introduce Timestral, a family of four time series foundation models for probabilistic forecasting, ranging from 5 million to 209 million parameters.
The authors developed Timestral models using a decoder-only transformer architecture trained with three main components: Decoupled Horizon Forecasting (DHF), which separates prediction tokens from the causally encoded context so that observed values serve as both forecast targets and context for later forecasts; a covariate framework that integrates known-future covariates via role-aware attention masks while preserving causality over target values; and a two-stage pipeline that first pretrains exclusively on a large synthetic corpus of univariate signals and structured multivariate dependencies, then post-trains on a mix of real and synthetic data with extended context.
How it works
before and afterThe Daily Edition
The front page in your inbox each weekday morning: the lead, the picks, and why each one earned its place.